watchdog: Fix sysctl consistency
[linux-2.6.git] / kernel / watchdog.c
1 /*
2  * Detect hard and soft lockups on a system
3  *
4  * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
5  *
6  * this code detects hard lockups: incidents in where on a CPU
7  * the kernel does not respond to anything except NMI.
8  *
9  * Note: Most of this code is borrowed heavily from softlockup.c,
10  * so thanks to Ingo for the initial implementation.
11  * Some chunks also taken from arch/x86/kernel/apic/nmi.c, thanks
12  * to those contributors as well.
13  */
14
15 #include <linux/mm.h>
16 #include <linux/cpu.h>
17 #include <linux/nmi.h>
18 #include <linux/init.h>
19 #include <linux/delay.h>
20 #include <linux/freezer.h>
21 #include <linux/kthread.h>
22 #include <linux/lockdep.h>
23 #include <linux/notifier.h>
24 #include <linux/module.h>
25 #include <linux/sysctl.h>
26
27 #include <asm/irq_regs.h>
28 #include <linux/perf_event.h>
29
30 int watchdog_enabled = 1;
31 int __read_mostly softlockup_thresh = 60;
32
33 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
34 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
35 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
36 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
37 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
38 #ifdef CONFIG_HARDLOCKUP_DETECTOR
39 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
40 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
41 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
42 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
43 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
44 #endif
45
46 /* boot commands */
47 /*
48  * Should we panic when a soft-lockup or hard-lockup occurs:
49  */
50 #ifdef CONFIG_HARDLOCKUP_DETECTOR
51 static int hardlockup_panic;
52
53 static int __init hardlockup_panic_setup(char *str)
54 {
55         if (!strncmp(str, "panic", 5))
56                 hardlockup_panic = 1;
57         else if (!strncmp(str, "0", 1))
58                 watchdog_enabled = 0;
59         return 1;
60 }
61 __setup("nmi_watchdog=", hardlockup_panic_setup);
62 #endif
63
64 unsigned int __read_mostly softlockup_panic =
65                         CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
66
67 static int __init softlockup_panic_setup(char *str)
68 {
69         softlockup_panic = simple_strtoul(str, NULL, 0);
70
71         return 1;
72 }
73 __setup("softlockup_panic=", softlockup_panic_setup);
74
75 static int __init nowatchdog_setup(char *str)
76 {
77         watchdog_enabled = 0;
78         return 1;
79 }
80 __setup("nowatchdog", nowatchdog_setup);
81
82 /* deprecated */
83 static int __init nosoftlockup_setup(char *str)
84 {
85         watchdog_enabled = 0;
86         return 1;
87 }
88 __setup("nosoftlockup", nosoftlockup_setup);
89 /*  */
90
91
92 /*
93  * Returns seconds, approximately.  We don't need nanosecond
94  * resolution, and we don't need to waste time with a big divide when
95  * 2^30ns == 1.074s.
96  */
97 static unsigned long get_timestamp(int this_cpu)
98 {
99         return cpu_clock(this_cpu) >> 30LL;  /* 2^30 ~= 10^9 */
100 }
101
102 static unsigned long get_sample_period(void)
103 {
104         /*
105          * convert softlockup_thresh from seconds to ns
106          * the divide by 5 is to give hrtimer 5 chances to
107          * increment before the hardlockup detector generates
108          * a warning
109          */
110         return softlockup_thresh / 5 * NSEC_PER_SEC;
111 }
112
113 /* Commands for resetting the watchdog */
114 static void __touch_watchdog(void)
115 {
116         int this_cpu = smp_processor_id();
117
118         __this_cpu_write(watchdog_touch_ts, get_timestamp(this_cpu));
119 }
120
121 void touch_softlockup_watchdog(void)
122 {
123         __this_cpu_write(watchdog_touch_ts, 0);
124 }
125 EXPORT_SYMBOL(touch_softlockup_watchdog);
126
127 void touch_all_softlockup_watchdogs(void)
128 {
129         int cpu;
130
131         /*
132          * this is done lockless
133          * do we care if a 0 races with a timestamp?
134          * all it means is the softlock check starts one cycle later
135          */
136         for_each_online_cpu(cpu)
137                 per_cpu(watchdog_touch_ts, cpu) = 0;
138 }
139
140 #ifdef CONFIG_HARDLOCKUP_DETECTOR
141 void touch_nmi_watchdog(void)
142 {
143         if (watchdog_enabled) {
144                 unsigned cpu;
145
146                 for_each_present_cpu(cpu) {
147                         if (per_cpu(watchdog_nmi_touch, cpu) != true)
148                                 per_cpu(watchdog_nmi_touch, cpu) = true;
149                 }
150         }
151         touch_softlockup_watchdog();
152 }
153 EXPORT_SYMBOL(touch_nmi_watchdog);
154
155 #endif
156
157 void touch_softlockup_watchdog_sync(void)
158 {
159         __raw_get_cpu_var(softlockup_touch_sync) = true;
160         __raw_get_cpu_var(watchdog_touch_ts) = 0;
161 }
162
163 #ifdef CONFIG_HARDLOCKUP_DETECTOR
164 /* watchdog detector functions */
165 static int is_hardlockup(void)
166 {
167         unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
168
169         if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
170                 return 1;
171
172         __this_cpu_write(hrtimer_interrupts_saved, hrint);
173         return 0;
174 }
175 #endif
176
177 static int is_softlockup(unsigned long touch_ts)
178 {
179         unsigned long now = get_timestamp(smp_processor_id());
180
181         /* Warn about unreasonable delays: */
182         if (time_after(now, touch_ts + softlockup_thresh))
183                 return now - touch_ts;
184
185         return 0;
186 }
187
188 #ifdef CONFIG_HARDLOCKUP_DETECTOR
189 static struct perf_event_attr wd_hw_attr = {
190         .type           = PERF_TYPE_HARDWARE,
191         .config         = PERF_COUNT_HW_CPU_CYCLES,
192         .size           = sizeof(struct perf_event_attr),
193         .pinned         = 1,
194         .disabled       = 1,
195 };
196
197 /* Callback function for perf event subsystem */
198 static void watchdog_overflow_callback(struct perf_event *event, int nmi,
199                  struct perf_sample_data *data,
200                  struct pt_regs *regs)
201 {
202         /* Ensure the watchdog never gets throttled */
203         event->hw.interrupts = 0;
204
205         if (__this_cpu_read(watchdog_nmi_touch) == true) {
206                 __this_cpu_write(watchdog_nmi_touch, false);
207                 return;
208         }
209
210         /* check for a hardlockup
211          * This is done by making sure our timer interrupt
212          * is incrementing.  The timer interrupt should have
213          * fired multiple times before we overflow'd.  If it hasn't
214          * then this is a good indication the cpu is stuck
215          */
216         if (is_hardlockup()) {
217                 int this_cpu = smp_processor_id();
218
219                 /* only print hardlockups once */
220                 if (__this_cpu_read(hard_watchdog_warn) == true)
221                         return;
222
223                 if (hardlockup_panic)
224                         panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
225                 else
226                         WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu);
227
228                 __this_cpu_write(hard_watchdog_warn, true);
229                 return;
230         }
231
232         __this_cpu_write(hard_watchdog_warn, false);
233         return;
234 }
235 static void watchdog_interrupt_count(void)
236 {
237         __this_cpu_inc(hrtimer_interrupts);
238 }
239 #else
240 static inline void watchdog_interrupt_count(void) { return; }
241 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
242
243 /* watchdog kicker functions */
244 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
245 {
246         unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
247         struct pt_regs *regs = get_irq_regs();
248         int duration;
249
250         /* kick the hardlockup detector */
251         watchdog_interrupt_count();
252
253         /* kick the softlockup detector */
254         wake_up_process(__this_cpu_read(softlockup_watchdog));
255
256         /* .. and repeat */
257         hrtimer_forward_now(hrtimer, ns_to_ktime(get_sample_period()));
258
259         if (touch_ts == 0) {
260                 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
261                         /*
262                          * If the time stamp was touched atomically
263                          * make sure the scheduler tick is up to date.
264                          */
265                         __this_cpu_write(softlockup_touch_sync, false);
266                         sched_clock_tick();
267                 }
268                 __touch_watchdog();
269                 return HRTIMER_RESTART;
270         }
271
272         /* check for a softlockup
273          * This is done by making sure a high priority task is
274          * being scheduled.  The task touches the watchdog to
275          * indicate it is getting cpu time.  If it hasn't then
276          * this is a good indication some task is hogging the cpu
277          */
278         duration = is_softlockup(touch_ts);
279         if (unlikely(duration)) {
280                 /* only warn once */
281                 if (__this_cpu_read(soft_watchdog_warn) == true)
282                         return HRTIMER_RESTART;
283
284                 printk(KERN_ERR "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
285                         smp_processor_id(), duration,
286                         current->comm, task_pid_nr(current));
287                 print_modules();
288                 print_irqtrace_events(current);
289                 if (regs)
290                         show_regs(regs);
291                 else
292                         dump_stack();
293
294                 if (softlockup_panic)
295                         panic("softlockup: hung tasks");
296                 __this_cpu_write(soft_watchdog_warn, true);
297         } else
298                 __this_cpu_write(soft_watchdog_warn, false);
299
300         return HRTIMER_RESTART;
301 }
302
303
304 /*
305  * The watchdog thread - touches the timestamp.
306  */
307 static int watchdog(void *unused)
308 {
309         static struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
310         struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
311
312         sched_setscheduler(current, SCHED_FIFO, &param);
313
314         /* initialize timestamp */
315         __touch_watchdog();
316
317         /* kick off the timer for the hardlockup detector */
318         /* done here because hrtimer_start can only pin to smp_processor_id() */
319         hrtimer_start(hrtimer, ns_to_ktime(get_sample_period()),
320                       HRTIMER_MODE_REL_PINNED);
321
322         set_current_state(TASK_INTERRUPTIBLE);
323         /*
324          * Run briefly once per second to reset the softlockup timestamp.
325          * If this gets delayed for more than 60 seconds then the
326          * debug-printout triggers in watchdog_timer_fn().
327          */
328         while (!kthread_should_stop()) {
329                 __touch_watchdog();
330                 schedule();
331
332                 if (kthread_should_stop())
333                         break;
334
335                 set_current_state(TASK_INTERRUPTIBLE);
336         }
337         __set_current_state(TASK_RUNNING);
338
339         return 0;
340 }
341
342
343 #ifdef CONFIG_HARDLOCKUP_DETECTOR
344 static int watchdog_nmi_enable(int cpu)
345 {
346         struct perf_event_attr *wd_attr;
347         struct perf_event *event = per_cpu(watchdog_ev, cpu);
348
349         /* is it already setup and enabled? */
350         if (event && event->state > PERF_EVENT_STATE_OFF)
351                 goto out;
352
353         /* it is setup but not enabled */
354         if (event != NULL)
355                 goto out_enable;
356
357         /* Try to register using hardware perf events */
358         wd_attr = &wd_hw_attr;
359         wd_attr->sample_period = hw_nmi_get_sample_period();
360         event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback);
361         if (!IS_ERR(event)) {
362                 printk(KERN_INFO "NMI watchdog enabled, takes one hw-pmu counter.\n");
363                 goto out_save;
364         }
365
366         printk(KERN_ERR "NMI watchdog disabled for cpu%i: unable to create perf event: %ld\n",
367                cpu, PTR_ERR(event));
368         return PTR_ERR(event);
369
370         /* success path */
371 out_save:
372         per_cpu(watchdog_ev, cpu) = event;
373 out_enable:
374         perf_event_enable(per_cpu(watchdog_ev, cpu));
375 out:
376         return 0;
377 }
378
379 static void watchdog_nmi_disable(int cpu)
380 {
381         struct perf_event *event = per_cpu(watchdog_ev, cpu);
382
383         if (event) {
384                 perf_event_disable(event);
385                 per_cpu(watchdog_ev, cpu) = NULL;
386
387                 /* should be in cleanup, but blocks oprofile */
388                 perf_event_release_kernel(event);
389         }
390         return;
391 }
392 #else
393 static int watchdog_nmi_enable(int cpu) { return 0; }
394 static void watchdog_nmi_disable(int cpu) { return; }
395 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
396
397 /* prepare/enable/disable routines */
398 static int watchdog_prepare_cpu(int cpu)
399 {
400         struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
401
402         WARN_ON(per_cpu(softlockup_watchdog, cpu));
403         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
404         hrtimer->function = watchdog_timer_fn;
405
406         return 0;
407 }
408
409 static int watchdog_enable(int cpu)
410 {
411         struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
412         int err;
413
414         /* enable the perf event */
415         err = watchdog_nmi_enable(cpu);
416         if (err)
417                 return err;
418
419         /* create the watchdog thread */
420         if (!p) {
421                 p = kthread_create(watchdog, (void *)(unsigned long)cpu, "watchdog/%d", cpu);
422                 if (IS_ERR(p)) {
423                         printk(KERN_ERR "softlockup watchdog for %i failed\n", cpu);
424                         return PTR_ERR(p);
425                 }
426                 kthread_bind(p, cpu);
427                 per_cpu(watchdog_touch_ts, cpu) = 0;
428                 per_cpu(softlockup_watchdog, cpu) = p;
429                 wake_up_process(p);
430         }
431
432         return 0;
433 }
434
435 static void watchdog_disable(int cpu)
436 {
437         struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
438         struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
439
440         /*
441          * cancel the timer first to stop incrementing the stats
442          * and waking up the kthread
443          */
444         hrtimer_cancel(hrtimer);
445
446         /* disable the perf event */
447         watchdog_nmi_disable(cpu);
448
449         /* stop the watchdog thread */
450         if (p) {
451                 per_cpu(softlockup_watchdog, cpu) = NULL;
452                 kthread_stop(p);
453         }
454 }
455
456 static void watchdog_enable_all_cpus(void)
457 {
458         int cpu;
459
460         watchdog_enabled = 0;
461
462         for_each_online_cpu(cpu)
463                 if (!watchdog_enable(cpu))
464                         /* if any cpu succeeds, watchdog is considered
465                            enabled for the system */
466                         watchdog_enabled = 1;
467
468         if (!watchdog_enabled)
469                 printk(KERN_ERR "watchdog: failed to be enabled on some cpus\n");
470
471 }
472
473 static void watchdog_disable_all_cpus(void)
474 {
475         int cpu;
476
477         for_each_online_cpu(cpu)
478                 watchdog_disable(cpu);
479
480         /* if all watchdogs are disabled, then they are disabled for the system */
481         watchdog_enabled = 0;
482 }
483
484
485 /* sysctl functions */
486 #ifdef CONFIG_SYSCTL
487 /*
488  * proc handler for /proc/sys/kernel/nmi_watchdog
489  */
490
491 int proc_dowatchdog_enabled(struct ctl_table *table, int write,
492                      void __user *buffer, size_t *length, loff_t *ppos)
493 {
494         proc_dointvec(table, write, buffer, length, ppos);
495
496         if (watchdog_enabled)
497                 watchdog_enable_all_cpus();
498         else
499                 watchdog_disable_all_cpus();
500         return 0;
501 }
502
503 int proc_dowatchdog_thresh(struct ctl_table *table, int write,
504                              void __user *buffer,
505                              size_t *lenp, loff_t *ppos)
506 {
507         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
508 }
509 #endif /* CONFIG_SYSCTL */
510
511
512 /*
513  * Create/destroy watchdog threads as CPUs come and go:
514  */
515 static int __cpuinit
516 cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
517 {
518         int hotcpu = (unsigned long)hcpu;
519         int err = 0;
520
521         switch (action) {
522         case CPU_UP_PREPARE:
523         case CPU_UP_PREPARE_FROZEN:
524                 err = watchdog_prepare_cpu(hotcpu);
525                 break;
526         case CPU_ONLINE:
527         case CPU_ONLINE_FROZEN:
528                 if (watchdog_enabled)
529                         err = watchdog_enable(hotcpu);
530                 break;
531 #ifdef CONFIG_HOTPLUG_CPU
532         case CPU_UP_CANCELED:
533         case CPU_UP_CANCELED_FROZEN:
534                 watchdog_disable(hotcpu);
535                 break;
536         case CPU_DEAD:
537         case CPU_DEAD_FROZEN:
538                 watchdog_disable(hotcpu);
539                 break;
540 #endif /* CONFIG_HOTPLUG_CPU */
541         }
542         return notifier_from_errno(err);
543 }
544
545 static struct notifier_block __cpuinitdata cpu_nfb = {
546         .notifier_call = cpu_callback
547 };
548
549 void __init lockup_detector_init(void)
550 {
551         void *cpu = (void *)(long)smp_processor_id();
552         int err;
553
554         err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
555         WARN_ON(notifier_to_errno(err));
556
557         cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
558         register_cpu_notifier(&cpu_nfb);
559
560         return;
561 }